Skill-Driven Data Sampling and Deep Learning Framework for Minute-Scale Solar Forecasting with Sky Images

Abstract

Accurate very short-term solar irradiance forecasting is crucial for optimizing the integration of solar energy into power systems. Herein, an image-based deep learning framework for minute-scale solar irradiance prediction is presented. The locally developed model is benchmarked against two commercial forecasting solutions deployed at the same experimental site, demonstrating superior accuracy and adaptability. A key contribution is the introduction of a skill-driven sampling algorithm based on clear sky index persistence error, which optimizes the training dataset by excluding low-utility samples while retaining essential physical features like solar zenith and azimuth angles. This algorithm enables the exclusion of up to 30% of the original training data, resulting in ≈16% savings in computational resources without affecting forecast accuracy validated using a test set of 324 991 observations. The model achieves a skill score of 7.63%, significantly outperforming the commercial models, which exhibit negative skill scores under the same conditions.